【作者】 丁晖；

【导师】 谢赤；

【作者基本信息】 湖南大学 ， 管理科学与工程， 2008， 博士

【摘要】 在开放经济中,汇率作为核心经济变量,调整并联系着各种宏观和微观经济因素,同时也影响着各经济体的内外均衡。人民币汇率是调节中国国民经济内外均衡的杠杆,也是维系中国与其他国家间经济往来的重要纽带和桥梁。特别是2005年7月21日汇率改革以后,人民币汇率问题已经成为影响世界经济形势以及中国与以美国和欧盟为代表的重要贸易伙伴之间经贸关系的关键问题。因此,探索汇率系统的内在规律,以更好地捕捉汇率行为的特征并进行准确的预测具有很强的理论意义与现实价值。汇率行为的复杂性使得对其有效预测的研究应基于非线性的范式之上。人工神经网络技术是一种非线性系统逼近和建模的有效工具,由于其具有良好的非线性映射能力及自适应、自学习和泛化能力,特别是包含反馈过程的动态神经网络能够更直接更生动地反映系统的动态特性,因而利用神经网络技术来对动态、非线性的汇率系统进行预测别具潜力。本文首先从汇率预测的基本理论、主要模型以及技术方法3方面介绍并评述了国内外汇率预测的研究现状;接着概述了神经网络技术的发展、特性及原理;在讨论了泛化能力的概念及在神经网络训练中最容易产生的过拟合问题之后,针对产生过拟合问题的两个主要原因,分别从神经网络模型本身和网络的训练过程两个方面,讨论了在汇率预测建模中一些关键参数的设计方法;最后讨论了几种预测效果评价指标。在实证研究部分,本文首先基于非线性理论从4个方面对汇率序列的非线性特征进行检验,发现人民币兑美元与人民币兑欧元汇率序列均具有较为复杂的非线性动态特征,适于采用神经网络这类非线性方法对其进行描述和预测。接着,本文以人民币兑美元汇率与人民币兑欧元汇率价格的水平和波动序列的数据特征为基础,从最优滞后期和最佳训练样本数等方面对影响神经网络模型预测能力的各关键参数进行了估计,并采用以MLP网络为代表的静态前馈网络和3种基本的动态反馈神经网络模型分别对几种汇率序列进行拟合与预测。通过对不同自由度下的各神经网络模型和简单随机游走模型的预测效果进行对比,发现在总体上动态反馈神经网络模型对人民币汇率各序列的样本内拟合及样本外预测能力,均优于静态前向神经网络和简单随机游走模型,且模型的样本内拟合效果与其样本外预测能力之间并无直接相关关系。具体的运算结果则表明,RNN2(1)和RNN3(1)模型分别为样本外前4周人民币兑美元汇率水平和第一周波动预测的最优模型,而RNN2(1)和RNN1(1)分别为样本外前4周人民币兑欧元汇率水平和第1周与第2周波动变化预测的最优模型,且其预测精度较其它模型存在显著优势,该结论支持了不同神经网络模型在汇率时间序列上的预测能力依赖于不同的确定汇率时间序列的假设。综观全文,本文无论在理论上还是在实证研究部分,都做出了一定创新。同时,本文的研究结果为准确地预测人民币汇率的价格水平和波动变化,并进一步为中央银行制定正确的外汇干预和货币政策、企业正确规避外汇风险等的决策均具有一定的实际指导作用。更多还原

【Abstract】 In an open economy, as a key variable of the economic systerm, exchange rate not only adjust and associate with various macro-and micro-economic factors, but also impact the balance of internal and external economies. RMB exchange rate system is the lever, which regulates internal and external balance of China’s national economy, and the important bridge to maintain economic ties between China and other countries as well. Especially after the exchange rate reform implemented from July 21, 2005, RMB exchange rate became the key issues which influences the world economic situation and trade relations between China and its important trading partners, such as the USA and the EU. Therefore, it has great theoretical significance and application value to explore the inherent laws of the system and capture its characteristic for accurately predicting the RMB exchange rate.The complexity of the exchange rate behavior calls for new normal formulas when making prediction study. The Artificial Neural Network technique is an effective tool for approximating and modeling nonlinear system. It has good nonlinear mapping capability, adaptive, self-learning and generalization ability. Especially, the dynamic neural network, which containing feedback process can directly and vividly reflect the dynamic nature of systems, thus it is attractive to use neural network technology to forecast the dynamic, non-linear exchange rate.In this paper, firstly we reviews and commentates the literatures of foreign exchange rate forecast, then introduces and analyzes from three perspectives, which are the basic theory of the exchange rate forecast, the main model and technical approach. Then a brief introduction was given on the artificial neural network technology, including its development, features and the theory. In the next part, first the concepts of generalization and the over-fitting issue in neural networks training process were introduced, then the author discusses some issues on the key parameters in modeling the neural network both from itself and the training process, according to the reason why over-fitting is caused. At last, some evaluation standards of forecasting performance were introduced.In the empirical studies part, the author first tests the non-linear features of the RMB exchange rate from four aspects, and finds that both the RMB/USD and the RMB/EUR exchange rate have complex nonlinear dynamic features. So the neural network, as one of the non-linear based approach, can be used for the fitting and forecasting. Then, from aspects such as optimal lag period and the best samples number of training collection, the author estimates the key parameters impacting the forecasting ability of neural network model. We adopted most used feed forward network MLP model and three kinds of basic dynamic feedback neural network model to train and forecast the time series of RMB exchange rate.By comparing the neural network models under different freedom parameters and the simple random walk model, we identify and select the optimal neural network model for each RMB exchange rate. The results showed in general that the recurrent network models is with beter performance in both in-sample fit and out-of-sample forcast on each RMB exchange rate time series. And for each ANN model, there is no direct relationship between the performance of in-sample fit and out-of-sample forcast.The accumulation in details proved that RNN2(1) and RNN3(1) are the best forecasting model respectively on the RMB/USD price level in four weeks and the fluctuations in the first week, RNN2(1) and RNN1(1) are the best forecasting model respectively on the RMB/EUR price level in four weeks and the fluctuations in the first week and the second week respectively. The overall predictive ability is superior to other models significantly. The conclusion proves the hypothesis that the prediction capability of different ANN model relies on different time series of exchange rate.In all, the paper has made some innovation both in the theoretical analysis and in the empirical application part. The study make an improvement in accurately predicting both the price level and the fluctuation changes of RMB exchange rate, which further have some practical usefulness for the Central Bank to formulate effective foreign exchange intervention and monetary policies, or for the corporate to avoid foreign exchange risk, etc.更多还原